Compute Library
 21.02
neon_cnn.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2016-2021 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
25 
26 #include "arm_compute/core/Types.h"
31 #include "utils/Utils.h"
32 
33 using namespace arm_compute;
34 using namespace utils;
35 
36 class NEONCNNExample : public Example
37 {
38 public:
39  bool do_setup(int argc, char **argv) override
40  {
41  ARM_COMPUTE_UNUSED(argc);
42  ARM_COMPUTE_UNUSED(argv);
43 
44  // Create memory manager components
45  // We need 2 memory managers: 1 for handling the tensors within the functions (mm_layers) and 1 for handling the input and output tensors of the functions (mm_transitions))
46  auto lifetime_mgr0 = std::make_shared<BlobLifetimeManager>(); // Create lifetime manager
47  auto lifetime_mgr1 = std::make_shared<BlobLifetimeManager>(); // Create lifetime manager
48  auto pool_mgr0 = std::make_shared<PoolManager>(); // Create pool manager
49  auto pool_mgr1 = std::make_shared<PoolManager>(); // Create pool manager
50  auto mm_layers = std::make_shared<MemoryManagerOnDemand>(lifetime_mgr0, pool_mgr0); // Create the memory manager
51  auto mm_transitions = std::make_shared<MemoryManagerOnDemand>(lifetime_mgr1, pool_mgr1); // Create the memory manager
52 
53  // The weights and biases tensors should be initialized with the values inferred with the training
54 
55  // Set memory manager where allowed to manage internal memory requirements
56  conv0 = std::make_unique<NEConvolutionLayer>(mm_layers);
57  conv1 = std::make_unique<NEConvolutionLayer>(mm_layers);
58  fc0 = std::make_unique<NEFullyConnectedLayer>(mm_layers);
59  softmax = std::make_unique<NESoftmaxLayer>(mm_layers);
60 
61  /* [Initialize tensors] */
62 
63  // Initialize src tensor
64  constexpr unsigned int width_src_image = 32;
65  constexpr unsigned int height_src_image = 32;
66  constexpr unsigned int ifm_src_img = 1;
67 
68  const TensorShape src_shape(width_src_image, height_src_image, ifm_src_img);
69  src.allocator()->init(TensorInfo(src_shape, 1, DataType::F32));
70 
71  // Initialize tensors of conv0
72  constexpr unsigned int kernel_x_conv0 = 5;
73  constexpr unsigned int kernel_y_conv0 = 5;
74  constexpr unsigned int ofm_conv0 = 8;
75 
76  const TensorShape weights_shape_conv0(kernel_x_conv0, kernel_y_conv0, src_shape.z(), ofm_conv0);
77  const TensorShape biases_shape_conv0(weights_shape_conv0[3]);
78  const TensorShape out_shape_conv0(src_shape.x(), src_shape.y(), weights_shape_conv0[3]);
79 
80  weights0.allocator()->init(TensorInfo(weights_shape_conv0, 1, DataType::F32));
81  biases0.allocator()->init(TensorInfo(biases_shape_conv0, 1, DataType::F32));
82  out_conv0.allocator()->init(TensorInfo(out_shape_conv0, 1, DataType::F32));
83 
84  // Initialize tensor of act0
85  out_act0.allocator()->init(TensorInfo(out_shape_conv0, 1, DataType::F32));
86 
87  // Initialize tensor of pool0
88  TensorShape out_shape_pool0 = out_shape_conv0;
89  out_shape_pool0.set(0, out_shape_pool0.x() / 2);
90  out_shape_pool0.set(1, out_shape_pool0.y() / 2);
91  out_pool0.allocator()->init(TensorInfo(out_shape_pool0, 1, DataType::F32));
92 
93  // Initialize tensors of conv1
94  constexpr unsigned int kernel_x_conv1 = 3;
95  constexpr unsigned int kernel_y_conv1 = 3;
96  constexpr unsigned int ofm_conv1 = 16;
97 
98  const TensorShape weights_shape_conv1(kernel_x_conv1, kernel_y_conv1, out_shape_pool0.z(), ofm_conv1);
99 
100  const TensorShape biases_shape_conv1(weights_shape_conv1[3]);
101  const TensorShape out_shape_conv1(out_shape_pool0.x(), out_shape_pool0.y(), weights_shape_conv1[3]);
102 
103  weights1.allocator()->init(TensorInfo(weights_shape_conv1, 1, DataType::F32));
104  biases1.allocator()->init(TensorInfo(biases_shape_conv1, 1, DataType::F32));
105  out_conv1.allocator()->init(TensorInfo(out_shape_conv1, 1, DataType::F32));
106 
107  // Initialize tensor of act1
108  out_act1.allocator()->init(TensorInfo(out_shape_conv1, 1, DataType::F32));
109 
110  // Initialize tensor of pool1
111  TensorShape out_shape_pool1 = out_shape_conv1;
112  out_shape_pool1.set(0, out_shape_pool1.x() / 2);
113  out_shape_pool1.set(1, out_shape_pool1.y() / 2);
114  out_pool1.allocator()->init(TensorInfo(out_shape_pool1, 1, DataType::F32));
115 
116  // Initialize tensor of fc0
117  constexpr unsigned int num_labels = 128;
118 
119  const TensorShape weights_shape_fc0(out_shape_pool1.x() * out_shape_pool1.y() * out_shape_pool1.z(), num_labels);
120  const TensorShape biases_shape_fc0(num_labels);
121  const TensorShape out_shape_fc0(num_labels);
122 
123  weights2.allocator()->init(TensorInfo(weights_shape_fc0, 1, DataType::F32));
124  biases2.allocator()->init(TensorInfo(biases_shape_fc0, 1, DataType::F32));
125  out_fc0.allocator()->init(TensorInfo(out_shape_fc0, 1, DataType::F32));
126 
127  // Initialize tensor of act2
128  out_act2.allocator()->init(TensorInfo(out_shape_fc0, 1, DataType::F32));
129 
130  // Initialize tensor of softmax
131  const TensorShape out_shape_softmax(out_shape_fc0.x());
132  out_softmax.allocator()->init(TensorInfo(out_shape_softmax, 1, DataType::F32));
133 
134  constexpr auto data_layout = DataLayout::NCHW;
135 
136  /* -----------------------End: [Initialize tensors] */
137 
138  /* [Configure functions] */
139 
140  // in:32x32x1: 5x5 convolution, 8 output features maps (OFM)
141  conv0->configure(&src, &weights0, &biases0, &out_conv0, PadStrideInfo(1 /* stride_x */, 1 /* stride_y */, 2 /* pad_x */, 2 /* pad_y */));
142 
143  // in:32x32x8, out:32x32x8, Activation function: relu
144  act0.configure(&out_conv0, &out_act0, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
145 
146  // in:32x32x8, out:16x16x8 (2x2 pooling), Pool type function: Max
147  pool0.configure(&out_act0, &out_pool0, PoolingLayerInfo(PoolingType::MAX, 2, data_layout, PadStrideInfo(2 /* stride_x */, 2 /* stride_y */)));
148 
149  // in:16x16x8: 3x3 convolution, 16 output features maps (OFM)
150  conv1->configure(&out_pool0, &weights1, &biases1, &out_conv1, PadStrideInfo(1 /* stride_x */, 1 /* stride_y */, 1 /* pad_x */, 1 /* pad_y */));
151 
152  // in:16x16x16, out:16x16x16, Activation function: relu
153  act1.configure(&out_conv1, &out_act1, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
154 
155  // in:16x16x16, out:8x8x16 (2x2 pooling), Pool type function: Average
156  pool1.configure(&out_act1, &out_pool1, PoolingLayerInfo(PoolingType::AVG, 2, data_layout, PadStrideInfo(2 /* stride_x */, 2 /* stride_y */)));
157 
158  // in:8x8x16, out:128
159  fc0->configure(&out_pool1, &weights2, &biases2, &out_fc0);
160 
161  // in:128, out:128, Activation function: relu
162  act2.configure(&out_fc0, &out_act2, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
163 
164  // in:128, out:128
165  softmax->configure(&out_act2, &out_softmax);
166 
167  /* -----------------------End: [Configure functions] */
168 
169  /*[ Add tensors to memory manager ]*/
170 
171  // We need 2 memory groups for handling the input and output
172  // We call explicitly allocate after manage() in order to avoid overlapping lifetimes
173  memory_group0 = std::make_unique<MemoryGroup>(mm_transitions);
174  memory_group1 = std::make_unique<MemoryGroup>(mm_transitions);
175 
176  memory_group0->manage(&out_conv0);
177  out_conv0.allocator()->allocate();
178  memory_group1->manage(&out_act0);
179  out_act0.allocator()->allocate();
180  memory_group0->manage(&out_pool0);
181  out_pool0.allocator()->allocate();
182  memory_group1->manage(&out_conv1);
183  out_conv1.allocator()->allocate();
184  memory_group0->manage(&out_act1);
185  out_act1.allocator()->allocate();
186  memory_group1->manage(&out_pool1);
187  out_pool1.allocator()->allocate();
188  memory_group0->manage(&out_fc0);
189  out_fc0.allocator()->allocate();
190  memory_group1->manage(&out_act2);
191  out_act2.allocator()->allocate();
192  memory_group0->manage(&out_softmax);
193  out_softmax.allocator()->allocate();
194 
195  /* -----------------------End: [ Add tensors to memory manager ] */
196 
197  /* [Allocate tensors] */
198 
199  // Now that the padding requirements are known we can allocate all tensors
200  src.allocator()->allocate();
201  weights0.allocator()->allocate();
202  weights1.allocator()->allocate();
203  weights2.allocator()->allocate();
204  biases0.allocator()->allocate();
205  biases1.allocator()->allocate();
206  biases2.allocator()->allocate();
207 
208  /* -----------------------End: [Allocate tensors] */
209 
210  // Populate the layers manager. (Validity checks, memory allocations etc)
211  mm_layers->populate(allocator, 1 /* num_pools */);
212 
213  // Populate the transitions manager. (Validity checks, memory allocations etc)
214  mm_transitions->populate(allocator, 2 /* num_pools */);
215 
216  return true;
217  }
218  void do_run() override
219  {
220  // Acquire memory for the memory groups
221  memory_group0->acquire();
222  memory_group1->acquire();
223 
224  conv0->run();
225  act0.run();
226  pool0.run();
227  conv1->run();
228  act1.run();
229  pool1.run();
230  fc0->run();
231  act2.run();
232  softmax->run();
233 
234  // Release memory
235  memory_group0->release();
236  memory_group1->release();
237  }
238 
239 private:
240  // The src tensor should contain the input image
241  Tensor src{};
242 
243  // Intermediate tensors used
244  Tensor weights0{};
245  Tensor weights1{};
246  Tensor weights2{};
247  Tensor biases0{};
248  Tensor biases1{};
249  Tensor biases2{};
250  Tensor out_conv0{};
251  Tensor out_conv1{};
252  Tensor out_act0{};
253  Tensor out_act1{};
254  Tensor out_act2{};
255  Tensor out_pool0{};
256  Tensor out_pool1{};
257  Tensor out_fc0{};
258  Tensor out_softmax{};
259 
260  // Neon allocator
262 
263  // Memory groups
264  std::unique_ptr<MemoryGroup> memory_group0{};
265  std::unique_ptr<MemoryGroup> memory_group1{};
266 
267  // Layers
268  std::unique_ptr<NEConvolutionLayer> conv0{};
269  std::unique_ptr<NEConvolutionLayer> conv1{};
270  std::unique_ptr<NEFullyConnectedLayer> fc0{};
271  std::unique_ptr<NESoftmaxLayer> softmax{};
272  NEPoolingLayer pool0{};
273  NEPoolingLayer pool1{};
274  NEActivationLayer act0{};
275  NEActivationLayer act1{};
276  NEActivationLayer act2{};
277 };
278 
279 /** Main program for cnn test
280  *
281  * The example implements the following CNN architecture:
282  *
283  * Input -> conv0:5x5 -> act0:relu -> pool:2x2 -> conv1:3x3 -> act1:relu -> pool:2x2 -> fc0 -> act2:relu -> softmax
284  *
285  * @param[in] argc Number of arguments
286  * @param[in] argv Arguments
287  */
288 int main(int argc, char **argv)
289 {
290  return utils::run_example<NEONCNNExample>(argc, argv);
291 }
Shape of a tensor.
Definition: TensorShape.h:39
int main(int argc, char **argv)
Main program for cnn test.
Definition: neon_cnn.cpp:288
1 channel, 1 F32 per channel
const DataLayout data_layout
Definition: Im2Col.cpp:151
Activation Layer Information class.
Definition: Types.h:1550
Includes all the Neon functions at once.
SimpleTensor< float > src
Definition: DFT.cpp:155
Copyright (c) 2017-2021 Arm Limited.
T x() const
Alias to access the size of the first dimension.
Definition: Dimensions.h:87
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
Definition: Error.h:152
T z() const
Alias to access the size of the third dimension.
Definition: Dimensions.h:97
Pooling Layer Information struct.
Definition: Types.h:1214
Abstract Example class.
Definition: Utils.h:78
Basic implementation of the tensor interface.
Definition: Tensor.h:37
Padding and stride information class.
Definition: Types.h:722
Num samples, channels, height, width.
Basic function to run cpu::kernels::CpuActivationKernel.
input allocator() -> allocate()
Basic function to simulate a pooling layer with the specified pooling operation.
Default malloc allocator implementation.
Definition: Allocator.h:36
Store the tensor&#39;s metadata.
Definition: TensorInfo.h:45
T y() const
Alias to access the size of the second dimension.
Definition: Dimensions.h:92
TensorShape & set(size_t dimension, size_t value, bool apply_dim_correction=true, bool increase_dim_unit=true)
Accessor to set the value of one of the dimensions.
Definition: TensorShape.h:79